Tech News - Data

Be it a tweet about taking the company private or smoking weed on a radio show, Elon Musk has been in news for all the wrong reasons recently and he is in news again but this time for what he is best admired as a modern day visionary. As per reports the Tesla and SpaceX founder is working on a 'superhuman' product that will connect your brain to a computer. We all know Musk along with eight others founded a company called Neuralink two years ago. The company has been developing implantable brain–computer interfaces, better known as BCIs. While in the short-term the company’s aim is to use the technology to treat brain diseases, Musk’s eventual goal is human enhancement, which he believes will make us more intelligent and powerful than even AI.  According to hints he gave a week ago, Neuralink may soon be close to announcing a product unlike anything we have seen: A brain computer interface. Appearing on the Joe Rogan Experience podcast last week, Musk stated that he’ll soon be announcing a new product – Neuralink – which will connect your brain to a computer, thus making you superhuman. When asked about Neuralink, Musk said "I think we’ll have something interesting to announce in a few months that’s better than anyone thinks is possible. Best case scenario, we effectively merge with AI. It will enable anyone who wants to have superhuman cognition. Anyone who wants. How much smarter are you with a phone or computer or without? You’re vastly smarter, actually. You can answer any question pretty much instantly. You can remember flawlessly. Your phone can remember videos [and] pictures perfectly. Your phone is already an extension of you. You’re already a cyborg. Most people don’t realise you’re already a cyborg. It’s just that the data rate, it’s slow, very slow. It’s like a tiny straw of information flow between your biological self and your digital self. We need to make that tiny straw like a giant river, a huge, high-bandwidth interface." If we visualize what Musk said, it feels like a scene straight from a Hollywood movie. However, many of the creations, from a decade ago, that were thought to belong solely in the world of science-fiction, have become a  reality now. Musk argues that through our over-dependence on smartphones, we have already taken the first step towards our cyborg future. Neuralink is an attempt to just accelerate the process by leaps and bounds. That's not all, Elon Musk was also quoted saying on CNBC. "If your biological self dies, you can upload into a new unit. Literally, with our Neuralink technology". Read the full news on CNBC. About Author Sandesh Deshpande is currently working as a System Administrator for Packt Publishing. He is highly interested in Artificial Intelligence and Machine Learning. Tesla is building its own AI hardware for self-driving cars Elon Musk’s tiny submarine is a lesson in how not to solve problems in tech. DeepMind, Elon Musk, and others pledge not to build lethal AI

In a paper published last month, Google developers explained the design, API, and implementation of TensorFlow.js, the JavaScript implementation of TensorFlow. TensorFlow.js was first introduced at the TensorFlow Dev Summit 2018. It is basically the successor of deeplearn.js, which was released in August 2017, and is now named as TensorFlow.js Core. Google’s motivation behind creating TensorFlow.js was to bring machine learning in the hands of web developers who generally do not have much experience with machine learning. It also aims at allowing experienced ML users and teaching enthusiasts to easily migrate their work to JS. The TensorFlow.js architecture TensorFlow.js, as the name suggests, is based on TensorFlow, with a few exceptions specific to the JS environment. This library comes with the following two sets of APIs: The Ops API facilitates lower-level linear algebra operations such as matrix, multiplication, tensor addition, and so on. The Layers API, similar to the Keras API, provide developers high-level model building blocks and best practices with emphasis on neural networks. Source: TensorFlow.js TensorFlow.js backends In order to support device-specific kernel implementations, TensorFlow.js has a concept of backends. Currently it supports three backends: the browser, WebGL, and Node.js. The two new rising web standards, WebAssembly and WebGPU, will also be supported as a backend by TensorFlow.js in the future. To utilize the GPU for fast parallelized computations, TensorFlow.js relies on WebGL, a cross-platform web standard that provides low-level 3D graphics APIs. Among the three TensorFlow.js backends, the WebGL backend has the highest complexity. With the introduction of Node.js and event-driven programming, the use of JS in server-side applications has grown over time. Server-side JS has full access to the filesystem, native operating system kernel, and existing C and C++ libraries. In order to support the server-side use cases of machine learning in JavaScript, TensorFlow.js comes with a Node.js backend that binds to the official TensorFlow C API using the N-API. As a fallback, TensorFlow.js provides a slower CPU implementation in plain JS. This fallback can run in any execution environment and is automatically used when the environment has no access to WebGL or the TensorFlow binary. Current applications of TensorFlow.js Since its launch, TensorFlow.js have seen its applications in various domains. Here are some of the interesting examples the paper lists: Gestural Interfaces TensorFlow.js is being used in applications that take gestural inputs with the help of webcam. Developers are using this library to build applications that translate sign language to speech translation, enable individuals with limited motor ability control a web browser with their face, and perform real-time facial recognition and pose-detection. Research dissemination The library has facilitated ML researchers to make their algorithms more accessible to others. For instance, the Magenta.js library, developed by the Magenta team, provides in-browser access to generative music models. Porting to the web with TensorFlow.js has increased the visibility of their work with their audience, namely musicians. Desktop and production applications In addition to web development, JavaScript has been used to develop desktop and production applications. Node Clinic, an open source performance profiling tool, recently integrated a TensorFlow.js model to separate CPU usage spikes by the user from those caused by Node.js internals. Another example is, Mood.gg Desktop, which is a desktop application powered by Electron, a popular JavaScript framework for writing cross-platform desktop apps. With the help of TensorFlow.js, Mood.gg detects which character the user is playing in the game called Overwatch, by looking at the user’s screen. It then plays a custom soundtrack from a music streaming site that matches with the playing style of the in-game character. Read the paper, Tensorflow.js: Machine Learning for the Web and Beyond, for more details. TensorFlow.js 0.11.1 releases! Emoji Scavenger Hunt showcases TensorFlow.js 16 JavaScript frameworks developers should learn in 2019

Plotly released Dash DAQ, a modern UI component library, which helps with data acquisition in Python, earlier this week. A data acquisition system (DAQ) helps collect, store, and distribute information. Dash DAQ is built on top of Plotly’s Dash (a Python framework used for building analytical web applications without requiring the use of JavaScript). Dash DAQ consists of 16 components. These components are used for building user interfaces that are capable of controlling and reading scientific instruments. To know more about each of their usage and configuration options, check out the official Dash DAQ components page. You can use Dash DAQ with Python drivers which are provided by instrument vendors. Alternatively, you can also write your own drivers with PySerial, PyUSB, or PyVISA. Dash DAQ is priced at $1980 as it is built with research labs in mind and is not suited currently for general python users. To install Dash DAQ, you have to purchase it first. After you make the purchase, a download page will automatically appear via which you can download it. Only one Dash DAQ library is allotted per developer. Here are the installation steps as mentioned in the official Dash DAQ installation page. Multiple apps of different variety have already been made using Dash DAQ. Here are some of the examples: Wireless Arduino Robot in Python, an app that wirelessly controls Sparki, an Arduino-based robot. Dash DAQ. Using Dash DAQ  for this app gives it clean, intuitive and virtual controls to build GUIs for your hardware. Robotic Arm in Python, an app that allows you to operate Robotic Arm Edge. Dash DAQ’s GUI components allow you to interface with all the robot’s motors and LED. Users can even do it via their mobile device, thereby enjoying the experience of a real remote control! Ocean Optics Spectrometer in Python, an app which allows users to interface with an Ocean Optics spectrometer. Here Dash DAQ offers interactive UI components which are written in Python allowing you to read and control the instrument in real-time. Apart from these few examples, there are a lot more applications that the developers at Plotly have built using Dash DAQ. plotly.py 3.0 releases 15 Useful Python Libraries to make your Data Science tasks Easier  

Yesterday, Facebook announced that Cadence, Esperanto, Intel, Marvell, and Qualcomm Technologies Inc, have committed to support their Glow compiler in future silicon products. Facebook, with this partnership aims to build a hardware ecosystem for machine learning. With Glow, their partners will be able to rapidly design and optimize new silicon products for AI and ML and help Facebook scale their platform. They are also planning to expand this ecosystem by adding more partners in 2018. What is Glow? Glow is a machine learning compiler which is used to speed up the performance of deep learning frameworks on different hardware platforms. The name “Glow” comes from Graph-Lowering, which is the main method that the compiler uses for generating efficient code. This compiler is designed to allow state-of-the-art compiler optimizations and code generation of neural network graphs. With Glow, hardware developers and researchers can focus on building next generation hardware accelerators that can be supported by deep learning frameworks like PyTorch. Hardware accelerators for ML solve a range of distinct problems. Some focus on inference, while others focus on training. How it works? Glow accepts a computation graph from deep learning frameworks such as, PyTorch and TensorFlow and generates highly optimized code for machine learning accelerators. To do so, it lowers the traditional neural network dataflow graph into a two-phase strongly-typed intermediate representation: Source: Facebook High-level intermediate representation allows the optimizer to perform domain-specific optimizations. Lower-level intermediate representation, an instruction-based address-only representation allows the compiler to perform memory-related optimizations, such as instruction scheduling, static memory allocation, and copy elimination. The optimizer then performs machine-specific code generation to take advantage of specialized hardware features. Glow supports a high number of input operators as well as a large number of hardware targets with the help of its lowering phase, which eliminates the need to implement all operators on all targets. The lowering phase reduces the input space and allows new hardware backends to focus on a small number of linear algebra primitives. You can read more about Facebook’s goals for Glow in its official announcement. If you are interesting in knowing how it works in more detail, check out this research paper and also its GitHub repository. Facebook launches LogDevice: An open source distributed data store designed for logs Google’s new What-if tool to analyze Machine Learning models and assess fairness without any coding Facebook introduces Rosetta, a scalable OCR system that understands text on images using Faster-RCNN and CNN

“Debugging code is drudgery” -Facebook Engineers Yue Jia, Ke Mao and Mark Harman To significantly reduce the amount of time developers spend on debugging code and rolling out new software, Facebook engineers have come up with an ingenious tool called ‘SapFix’. Sapfix, which is still under development, can automatically generate fixes for specific bugs  identified by Sapienz. It will then propose these fixes to engineers for approval and deployment to production. SapFix will eventually be able to operate independently from Sapienz, Facebook’s intelligent automated software testing tool. For now, it is a proof-of-concept that relies on the latter tool to pinpoint bugs. How does SapFix work? This AI hybrid tool will generate bug fixes depending upon the type of bug encountered. For instance: For simpler bugs: SapFix will create patches that revert the code submission that introduced these bugs. For complicated bugs: The tool uses a collection of “templated fixes” that were created by human engineers based on previous bug fixes. If human-designed template fixes aren’t up to the job: The tool attempts a “mutation-based fix,” which works by continuously making small modifications to the code that caused the software to crash, until a solution is found. SapFix generates multiple potential fixes for every bug. This is then submitted to the engineers for evaluation. The fixes are tested in advance so engineers can check if they might cause problems like compilation errors and other crashes. Source: Facebook With an automated end-to-end testing and repair, SapFix is an important milestone in AI hybrid tool deployment. Facebook intends to open source both, SapFix and Sapienz, once additional engineering work has been completed. You can read more about this tool on Facebook’s Blog. Facebook introduces Rosetta, a scalable OCR system that understands text on images using Faster-RCNN and CNN How AI is going to transform the Data Center Facebook Reality Labs launch SUMO Challenge to improve 3D scene understanding and modeling algorithms  

Last month, Nathan Egge, a Senior Research Engineer at Mozilla explained technical details behind AV1 in depth at the Mile High Video Workshop in Denver. AV1 is a new open source royalty-free video codec that promises to help companies and individuals transmit high-quality video over the internet efficiently. AV1 is developed by the Alliance for Open Media (AOMedia), an association of firms from the semiconductor industry, video on demand providers, and web browser developers, founded in 2015. Mozilla joined AOMedia as a founding member. AV1 was created for a broad set of industry use cases such as video on demand/streaming, video conferencing, screen sharing, video game streaming, and broadcast. It is widely supported and adopted and gives at least 30% better than current generation video codecs. The alliance was able to hit a key milestone with the release of AV1 1.0.0 specification earlier this year in June. The codec has seen increasing interest from various companies, for instance, YouTube launched AV1 Beta Playlist in September. The following diagram shows the various stages in the working of a video codec: Source: YouTube We will cover the tools and algorithm used in some of these stages. Let’s see some of its technical details from Egge’s talk: AV1 Profiles Profiles specify the bit depth and subsampling formats supported. In AV1 there are three profiles: Main, High, and Professional which differ in terms of their bit-depth and chroma subsampling. The following table shows their bit-depth and chroma subsampling: Main High Professional Bit depth 8-bit and 10-bit 8-bit and 10-bit 8-bit, 10-bit, and 12-bit Chroma subsampling 4:0:0, 4:2:0 4:0:0, 4:2:0, and 4:4:4 4:0:0, 4:2:0, 4:2:2, and 4:4:4 High-level syntax In VP9 there is a concept of superframes that at some point becomes complicated. Superframes allows you to consolidate multiple coded frames into one single chunk. AV1 comes with high-level syntax that includes: sequence header, frame header, tile group, and tiles. Sequence header starts a video stream, frame headers are at the beginning of a frame, a tile group is an independent group of tiles, and finally, we have tiles which can be independently decoded. Source: YouTube Multi-symbol entropy coder Unlike VP9, which uses a tree-based boolean non-adaptive binary arithmetic encoder to encode all syntax elements, AV1 uses a symbol-to-symbol adaptive multi-symbol arithmetic coder. Each of its syntax element is a member of a specific alphabet of N elements, and a context is a set of N probabilities together with a count to facilitate fast early adaptation. Transform types In addition to DCT and ADST transform types, AV1 introduces two other transforms called flipped ADST and identity transform as extended transform types. Identity transform enables you to effectively code residual blocks with edges and lines. AV1 thus comes with the advantage of a total of sixteen horizontal and vertical transform type combinations. Intra prediction modes Along with the 8 main directional modes from VP9, up to 56 more directions are added but not all of them are available at smaller sizes. The following are some of the prediction modes introduced in AV1: Smooth H + V modes allow you to smoothly interpolate between values in the left column and last value in the above row. Palette mode is introduced to the intra coder as a general extra coding tool. It will be especially useful for artificial videos like screen capture and games, where blocks can be approximated by a small number of unique colors. The palette predictor for each plane of a block is depicted by: A color palette, with 2 to 8 colors Color indices for all pixels in the block Chroma from Luma (CfL) is a chroma-only intra predictor that models chroma pixels as a linear function of coincident reconstructed luma pixels. Source: YouTube First, the reconstructed luma pixels are subsampled into the chroma resolution, and then the DC component is removed to form the AC contribution. In order to approximate chroma AC component from the AC contribution, instead of requiring the decoder to imply scaling parameters, CfL determines the parameters based on the original chroma pixels and signals them in the bitstream. As a result, this reduces decoder complexity and yields more precise predictions. As for the DC prediction, it is computed using intra DC mode, which is sufficient for most chroma content and has mature fast implementations. Constrained Directional Enhancement Filter (CDEF) CDEF is a detail-preserving deringing filter, which is designed to be applied after deblocking. It works by estimating edge directions followed by applying a non-separable non-linear low-pass directional filter of size 5×5 with 12 non-zero weights. In order to avoid extra signaling, the decoder uses a normative fast search algorithm to compute the direction per 8×8 block that minimizes the quadratic error from a perfect directional pattern. Film Grain Synthesis In AV1, film grain synthesis is a normative post-processing applied outside of the encoding/decoding loop. Film grain is abundant in TV and movie content, which needs to be preserved while encoding. But, its random nature makes it difficult to compress with traditional coding tools. In film grain synthesis, the grain is removed from the content before compression, its parameters are estimated and then sent in the AV1 bitstream. The grain is then synthesized based on the received parameters and added to the reconstructed video. For grainy content, film grain synthesis significantly reduces the bitrate necessary to reconstruct the grain with sufficient quality. You can watch Into the Depths The Technical Details behind AV1 by Nathan Egge on YouTube: https://www.youtube.com/watch?v=On9VOnIBSEs&t=463s Presenting dav1d, a new lightweight AV1 decoder, by VideoLAN and FFmpeg YouTube starts testing AV1 video codec format, launches AV1 Beta Playlist Opus 1.3, a popular FOSS audio codec with machine learning and VR support, is now generally available

Today Baidu released a continual natural language processing framework ERNIE 2.0. ERNIE stands for Enhanced Representation through kNowledge IntEgration. Baidu claims in its research paper that ERNIE 2.0 outperforms BERT and the recent XLNet in 16 NLP tasks in Chinese and English. Additionally, Baidu has open sourced ERNIE 2.0 model. In March Baidu had announced the release of ERNIE 1.0, its pre-trained model based on PaddlePaddle, Baidu’s deep learning open platform. According to Baidu, ERNIE 1.0 outperformed BERT in all Chinese language understanding tasks. Pre-training procedures of the models such as BERT, XLNet and ERNIE 1.0 are mainly based on a few simple tasks modeling co-occurrence of words or sentences, highlights the paper. For example, BERT constructed a bidirectional language model task and the next sentence prediction task to capture the co-occurrence information of words and sentences; XLNet constructed a permutation language model task to capture the co-occurrence information of words. But besides co-occurring information, there are much richer lexical, syntactic and semantic information in training corpora. For example, named entities, such as person names, place names, and organization names, contain concept information; Sentence order and sentence proximity information can enable the models to learn structure-aware representations; Semantic similarity at the document level or discourse relations among sentences can enable the models to learn semantic-aware representations. So is it possible to further improve the performance if the model was trained to learn more kinds of tasks constantly? Source: ERNIE 2.0 research paper Based on this idea, Baidu has proposed a continual pre-training framework for language understanding in which pre-training tasks can be incrementally built and learned through multi-task learning in a continual way. According to Baidu, in this framework, different customized tasks can be incrementally introduced at any time and these tasks are trained through multi-task learning, which enables the encoding of lexical, syntactic and semantic information across tasks. And whenever a new task arrives, this framework can incrementally train the distributed representations without forgetting the previously trained parameters. The Structure of Released ERNIE 2.0 Model Source: ERNIE 2.0 research paper ERNIE is a continual pre-training framework which provides a feasible scheme for developers to build their own NLP models. The fine-tuning source codes of ERNIE 2.0 and pre-trained English version models can be downloaded from the GitHub page. The team at Baidu compared the performance of ERNIE 2.0 model with the existing  pre-training models on the English dataset GLUE and 9 popular Chinese datasets separately. The results show that ERNIE 2.0 model outperforms BERT and XLNet on 7 GLUE language understanding tasks and outperforms BERT on all of the 9 Chinese NLP tasks, such as DuReader Machine Reading Comprehension, Sentiment Analysis and Question Answering.  Specifically, according to the experimental results on GLUE datasets, ERNIE 2.0 model almost comprehensively outperforms BERT and XLNET on English tasks, whether it is a base model or the large model. Furthermore, the research paper shows that ERNIE 2.0 large model achieves the best performance and creates new results on the Chinese NLP tasks. Source: ERNIE 2.0 research paper To know more about ERNIE 2.0, read the research paper and check out their official blog on Baidu’s website. DeepMind’s AI uses reinforcement learning to defeat humans in multiplayer games CMU and Google researchers present XLNet: a new pre-training method for language modeling that outperforms BERT on 20 tasks Transformer-XL: A Google architecture with 80% longer dependency than RNNs  

After releasing PostgreSQL 11 beta 1, back in May, the PostgreSQL Global Development Group finally released PostgreSQL 11, yesterday. PostgreSQL 11 explores features such as increased performance for partitioning, support for transactions in stored procedures, improved capabilities for query parallelism, and Just-in-Time (JIT) compilation for expressions among other updates. PostgreSQL is a popular open source relational database management system that offers better reliability, robustness, and enhanced performance measures. Let’s have a look at these features in PostgreSQL 11. Increased performance for partitioning PostgreSQL 11 comes with an ability to partition the data using a hash key, which is known as hash partitioning. This adds to the already existing ability to partition data in PostgreSQL using a list of values or by a range. Moreover, PostgreSQL 11 also improves the data federation abilities by implementing functionality improvements for partitions using PostgreSQL foreign data wrapper, and postgres_fdw. For managing these partitions, PostgreSQL 11 comes with a “catch-all” default partition for data that doesn’t match a partition key. It also comes with an ability to create primary keys, foreign keys, indexes as well as triggers on partitioned tables. The latest release also offers support for automatic movement of rows to the correct partition, given that the partition key for that row is updated. Additionally, PostgreSQL 11 enhances the query performance when reading from partitions with the help of a new partition elimination strategy. It also offers support for the popular "upsert" feature on partitioned tables. The upsert feature helps users simplify the application code as well as reduce the network overhead when interacting with their data. Support for transactions in stored procedures With PostgreSQL 11 comes newly added SQL procedures that help perform full transaction management within the body of a function. This enables the developers to build advanced server-side applications like the ones that involve incremental bulk data loading. Also, SQL procedures can now be created using the CREATE PROCEDURE command which is executed using the CALL command. These SQL procedures are supported by the server-side procedural languages such as PL/pgSQL, PL/Perl, PL/Python, and PL/Tcl. Improved capabilities for query parallelism PostgreSQL 11 enhances the parallel query performance, using the performance gains in parallel sequential scans and hash joins. It also performs more efficient scans of the partitioned data. PostgreSQL 11 comes with added parallelism for a range of data definitions commands, especially for the creation of B-tree indexes generated by executing the standard CREATE INDEX command. Other data definition commands that either create tables or materialize the views from queries are also enabled with parallelism. This includes the CREATE TABLE .. AS, SELECT INTO, and CREATE MATERIALIZED VIEW. Just-in-Time (JIT) compilation for expressions PostgreSQL 11 offers support for Just-In-Time (JIT) compilation, This helps to accelerate the execution of certain expressions during query execution. The JIT expression compilation uses the LLVM project to boost the execution of expressions in WHERE clauses, target lists, aggregates, projections, as well as some other internal operations. Other Improvements ALTER TABLE .. ADD COLUMN .. DEFAULT ..have been replaced with a not NULL default to rewrite the whole table on execution. This offers a significant performance boost when running this command. Additional functionality has been added for working with window functions, including allowing RANGE to use PRECEDING/FOLLOWING, GROUPS, and frame exclusion. Keywords such as "quit" and "exit" have been added to the PostgreSQL command-line interface to help make it easier to leave the command-line tool. For more information, check out the official release notes. PostgreSQL group releases an update to 9.6.10, 9.5.14, 9.4.19, 9.3.24 How to perform data partitioning in PostgreSQL 10 How to write effective Stored Procedures in PostgreSQL

Drivendata has come out with a new tool, named, Deon, which allows you to easily add an ethics checklist to your data science projects. Deon is aimed at pushing the conversation about ethics in data science, machine learning, and Artificial intelligence by providing actionable reminders to data scientists. According to the Deon team, “it's not up to data scientists alone to decide what the ethical course of action is. This has always been a responsibility of organizations that are part of civil society. This checklist is designed to provoke conversations around issues where data scientists have particular responsibility and perspective”. Deon comes with a default checklist, but you can also develop your own custom checklists by removing items and sections, or marking items as N/A depending on the needs of the project. There are also real-world examples linked with each item in the default checklist.   To be able to run Deon for your data science projects, you need to have Python 3 or greater. Let’s now discuss the two types of checklists, Default, and Custom, that comes with Deon. Default checklist The default checklist comprises of sections on Data Collection, Data Storage, Analysis, Modeling, and Deployment. Data Collection This checklist covers information on informed consent, Collection Bias, and Limit PII exposure. Informed consent includes a mechanism for gathering consent where users have clear understanding of what they are consenting to. Collection Bias checks on sources of bias introduced during data collection and survey design. Lastly, Limit PII exposure talks about ways that can help minimize the exposure of personally identifiable information (PII). Data Storage This checklist covers sections such as Data security, Right to be forgotten and Data retention plan. Data Security refers to a plan to protect and secure data. Right to be forgotten includes a mechanism by which an individual can have his/her personal information. Data Retention consists of a plan to delete the data if no longer needed. Analysis This section comprises information on Missing perspectives, Dataset bias, Honest representation, Privacy in analysis and Auditability. Missing perspectives address the blind spots in data analysis via engagement with relevant stakeholders. Dataset bias discusses examining the data for possible sources of bias and consists of steps to mitigate or address them. Honest representation checks if visualizations, summary statistics, and reports designed honestly represent the underlying data. Privacy in analysis ensures that the data with PII are not used or displayed unless necessary for the analysis. Auditability refers to the process of producing an analysis which is well documented and reproducible. Modeling This offers information on Proxy discrimination, Fairness across groups, Metric selection,  Explainability, and Communicate bias. Proxy discrimination talks about ensuring that the model does not rely on variables or proxies that are discriminatory. Fairness across groups is a section that cross-checks whether the model results have been tested for fairness with respect to different affected groups. Metric selection considers the effects of optimizing for defined metrics and other additional metrics. Explainability talks about explaining the model’s decision in understandable terms. Communicate bias makes sure that the shortcomings, limitations, and biases of the model have been properly communicated to relevant stakeholders. Deployment This covers topics such as Redress, Roll back, Concept drift, and Unintended use. Redress discusses with an organization a plan for response in case users get harmed by the results. Roll back talks about a way to turn off or roll back the model in production when required. Concept drift refers to changing relationships between input and output data in a problem over time. This part in a checklist reminds the user to test and monitor the concept drift. This is to ensure that the model remains fair over time. Unintended use prompts the user about the steps to be taken for identifying and preventing uses and abuse of the model. Custom checklists For your projects with particular concerns, it is recommended to create your own checklist.yml file. Custom checklists are required to follow the same schema as checklist.yml. Custom Checklists need to have a top-level title which is a string, and sections which are a list. Each section in the list must have a title, a section_id, and then a list of lines. Each line must include a line_id, a line_summary, and a line string which is the content. When changing the default checklist, it is necessary to keep in mind that Deon’s goal is to have checklist items that are actionable. This is why users are advised to avoid suggesting items that are vague (e.g., "do no harm") or extremely specific (e.g., "remove social security numbers from data"). For more information, be sure to check out the official Drivendata blog post. The Cambridge Analytica scandal and ethics in data science OpenAI charter puts safety, standards, and transparency first 20 lessons on bias in machine learning systems by Kate Crawford at NIPS 2017

In a paper published this month, the human motions to wear clothes is synthesized in animation with reinforcement learning. The paper named Learning to Dress: Synthesizing Human Dressing Motion via Deep Reinforcement Learning was published yesterday. The team is made up of two Ph.D. students from The Georgia Institute of Technology, two of its professors and a researcher from Google Brain. Understanding the dressing problem Dressing, putting on a t-shirt or a jacket is something we do every day. Yet it is a computationally costly and complex task for a machine to perform or be simulated by computers. Techniques in physics simulation and machine learning are used in this paper to simulate an animation. A physics engine is used to simulate character motion and cloth motion. On the other hand deep reinforcement learning on a neural network is used to produce character motion. Physics engine and reinforcement learning on a neural network The authors of the paper introduce a salient representation of haptic information to guide the dressing process. Then the haptic information is used in the reward function to provide learning signals when training the network. As the task is too complex to do perform in one go, the dressing task is separated into several subtasks for better control. A policy sequencing algorithm is introduced to match the distribution of output states from one task to the input distribution for the next task. The same approach is used to produce character controllers for various dressing tasks like wearing a t-shirt, wearing a jacket, and robot-assisted dressing of a sleeve. Dressing is complex, split into several subtasks The approach taken by the authors splits the dressing task into a sequence of subtasks. Then a state machine guides the between these tasks. Dressing a jacket, for example, consists of four subtasks: Pulling the sleeve over the first arm. Moving the second arm behind the back to get in position for the second sleeve. Putting hand in the second sleeve. Finally, returning the body back to a rest position. A separate reinforcement learning problem is formulated for each subtask in order to learn a control policy. The policy sequencing algorithm ensures that these individual control policies can lead to a successful dressing sequence on being executed sequentially. The algorithm matches the initial state of one subtask with the final state of the previous subtask in the sequence. A variety of successful dressing motions can be produced by applying the resultant control policies. Each subtask in the dressing task is formulated as a partially observable Markov Decision Process (POMDP). Character dynamics are simulated with Dynamic Animation and Robotics Toolkit (DART) and cloth dynamics with NVIDIA PhysX. Conclusion and room for improvement A system that learns to animate a character that puts on clothing is successfully created with the use of deep reinforcement learning and physics simulation. From the subtasks, the system learns each sub-task individually, then connects them with a state machine. It was found that carefully selecting the cloth observations and the reward functions were important factors for the success of the approach taken. This system currently performs only upper body dressing. For lower body, a balance into the controller would be required. The number of subtasks might reduce on using a control policy architecture with memory. This will allow for greater generalization of the skills learned. You can read the research paper at the Georgia Institute of Technology website. Facebook launches Horizon, its first open source reinforcement learning platform for large-scale products and services Deep reinforcement learning – trick or treat? Google open sources Active Question Answering (ActiveQA), a Reinforcement Learning based Q&A system

PyTorch has announced a preview of the blueprint for PyTorch 1.0, the next major release of the framework. This breakthrough version is expected to bring more stability, integration support and complete production backing allowing developers to move from core research to production in an amicable way without having to deal with any migration challenges. PyTorch is an open-source Python-based scientific computing package which provides powerful GPU acceleration. PyTorch is known for advanced indexing and functions, imperative style, integration support and API simplicity. This is one of the key reasons why developers prefer PyTorch for research and hackability. To know more about how Facebook-based PyTorch competes with Google’s TensorFlow read our take on this deep learning war. Some of the noteworthy changes in the roadmap for PyTorch 1.0 are: Production support One of the biggest challenges faced by developers in terms of using PyTorch is production support. There are n number of issues faced while trying to run the models efficiently in production environments. Even though PyTorch provides excellent simplicity and flexibility, due to its tight coupling to Python, the performance at production-scale is a challenge.   To counter these challenges, the PyTorch team has decided to bring PyTorch and Caffe2 together to provide production-scale readiness to the developers. However, adding production support brings complexity and configurable options for models in the API. The PyTorch team will stick to the goal of keeping the platform -- a favorable choice -- for researchers and developers. Hence, they are introducing a new just-in-time (JIT) compiler, named torch.jit. torch.jit compiler rewrites PyTorch models during runtime in order to achieve scalability and efficiency in production environments. It can also export PyTorch models to run in a C++ environment. (runtime based on Caffe2 bits) Note: In PyTorch version 1.0, your existing code will continue to work as-is. Let’s go through how JIT compiler can be used to export models to a Python-less environment in order to improve their working performance. torch.jit: The go-to compiler for your PyTorch models Building models using Python code, no doubt gives maximum productivity and makes PyTorch very simple and easy-to-use. However, this also means PyTorch finding it difficult to know which operation you will run next. This can be frustrating for the developers during model export and automatic performance optimizations because they need to be aware of how the computations will look like before it even gets implemented. To deal with these issues, PyTorch provides two ways of recovering information from the Python code. Both these methods will be useful based on different contexts, giving you the leverage to use/mix them with ease. Tracing the native Python code Compiling a subset of the Python language Tracing mode torch.jit.trace function allows you to record the native PyTorch operations performed along with the data dependencies between them. PyTorch version 0.3 already had a tracer function which is used to export models through ONNX. This new version uses a high-performance C++ runtime that allows PyTorch to re-execute programs for you. The key advantage of using this method is that it doesn’t have to deal with how your Python code is structured since we only trace through native PyTorch operations. Script mode PyTorch team has come up with a solution called scripting mode made specially for those models such as RNNs which make use of control flow. However, you will have to write out a regular Python function (avoiding complex language features) In order to get your function compiled, you can assign @script decorator. This will make sure it alters your Python function directly into high-performance C++ during runtime. Advantages in optimization and export techniques Irrespective of you using a trace or a script function, the technique allows you to optimize/export the model for use in production environments (i.e. Python-free portrayal of the model) Now you can derive bigger segments of the model into an intermediate representation to work with sophisticated models. You can use high-performance backends available in Caffe2 to run the models efficiently Usability If you don’t need to export or optimize your model, you do not need to use these set of new features. These modes will be included into the core of the PyTorch ecosystem, thus allowing you to mix and match them with the existing code seamlessly as per your needs. Additional changes and improvements In addition to the major update in the production support for 1.0, PyTorch team will continue working on optimizing, working on the stability of the interface, and fixing other modules in PyTorch ecosystem PyTorch 1.0 will see some changes in the backend side which might affect user-written C and C++ extensions. In order to incorporate new features and optimization techniques from Caffe2, PyTorch team is replacing (optimizing) the backend ATen library. PyTorch team is planning to release 1.0 during the summer. For the detailed preview of the roadmap, you can refer the official PyTorch blog. Top 10 deep learning frameworks The Deep Learning Framework Showdown: TensorFlow vs CNTK Why you should use Keras for deep learning

Yesterday, the Netflix team announced to open-source Metaflow, a Python library that helps scientists and engineers build and manage real-life data science projects. The Netflix team writes, “Over the past two years, Metaflow has been used internally at Netflix to build and manage hundreds of data-science projects from natural language processing to operations research.” Metaflow was developed by Netflix to boost productivity of data scientists who work on a wide variety of projects from classical statistics to deep learning. It provides a unified API to the infrastructure stack required to execute data science projects, from prototype to production. Metaflow integrates with Netflix's data science infrastructure stack Models are only a small part of an end-to-end data science project. Production-grade projects rely on a thick stack of infrastructure. At the minimum, projects need data and a way to perform computation on it. In a business environment like Netflix's typical data science project, the team touches upon all the layers of the stack depicted below: Source: Netflix website Data is accessed from a data warehouse, which can be a folder of files, a database, or a multi-petabyte data lake. The modeling code crunches the data executed in a compute environment and a job scheduler is used to orchestrate multiple units of work. Then the team architects the code to be executed by structuring it as an object hierarchy, Python modules, or packages. They version the code, input data, and produce ML models. After the model has been deployed to production, the team faces pertinent questions about model operations for example; How to keep the code running reliably in production? How to monitor its performance? How to deploy new versions of the code to run in parallel with the previous version? Additionally at the very top of the stack there are other questions like how to produce features for your models, or how to develop models in the first place using off-the-shelf libraries. In this Metaflow provides a unified approach to navigating the stack. Metaflow is more prescriptive about the lower levels of the stack but it is less opinionated about the actual data science at the top of the stack. Developers can use Metaflow with their favorite machine learning or data science libraries, such as PyTorch, Tensorflow, or  SciKit Learn. Metaflow allows you to write models and business logic as idiomatic Python code. Internally, Metaflow leverages existing infrastructure when feasible. The core value proposition of Metaflow is its integrated full-stack, human-centric API, rather than reinventing the stack itself. Metaflow on Amazon Web Services Metaflow is a cloud-native framework which it leverages elasticity of the cloud by design — both for compute and storage. Netflix is one of the largest users of Amazon Web Services (AWS) and have accumulated plenty of operational experience and expertise in dealing with the cloud. For this open-source release, Netflix partnered with AWS to provide a seamless integration between Metaflow and various AWS services. Metaflow comes with built-in capability to snapshot all code and data in Amazon S3 automatically, a key value proposition for the internal Metaflow setup. This provides data science teams with a comprehensive solution for versioning and experiment tracking without any user intervention, core of any production-grade machine learning infrastructure. In addition, Metaflow comes bundled with a high-performance S3 client, which can load data up to 10Gbps. Additionally Metaflow provides a first-class local development experience. It allows data scientists to develop and test code quickly on laptops, similar to any Python script. If the workflow supports parallelism, Metaflow takes advantage of all CPU cores available on the development machine. How is Metaflow different from existing Python frameworks On Hacker News, developers discuss how Metaflow is different than existing tools or workflows. One of them comments, “I don't like to criticise new frameworks / tools without first understanding them, but I like to know what some key differences are without the marketing/PR fluff before giving one a go. For instance, this tutorial example here does not look substantially different to what I could achieve just as easily in R, or other Python data wrangling frameworks. Is the main feature the fact I can quickly put my workflows into the cloud?” Someone from the Metaflow team responds on this thread, “Here are some key features: - Metaflow snapshots your code, data, and dependencies automatically in a content-addressed datastore, which is typically backed by S3, although local filesystem is supported too. This allows you to resume workflows, reproduce past results, and inspect anything about the workflow e.g. in a notebook. This is a core feature of Metaflow. - Metaflow is designed to work well with a cloud backend. We support AWS today but technically other clouds could be supported too. There's quite a bit of engineering that has gone into building this integration. For instance, using the Metaflow's built-in S3 client, you can pull over 10Gbps, which is more than you can get with e.g. aws CLI today easily. - We have spent time and effort in keeping the API surface area clean and highly usable. YMMV but it has been an appealing feature to many users this far.” Developers can find the project home page here and its code at GitHub. Netflix open sources Polynote, an IDE-like polyglot notebook with Scala support, Apache Spark integration, multi-language interoperability, and more Tesla Software Version 10.0 adds Smart Summon, in-car karaoke, Netflix, Hulu, and Spotify streaming Netflix security engineers report several TCP networking vulnerabilities in FreeBSD and Linux kernels  

Facebook has open-sourced MUSE, Multilingual Unsupervised and Supervised Embeddings. It is a Python library to align embedding spaces in a supervised or unsupervised way. The supervised method uses a bilingual dictionary or identical character strings. The unsupervised approach does not use any parallel data. Instead, it builds a bilingual dictionary between two languages by aligning monolingual word embedding spaces in an unsupervised way. Facebook MUSE has state-of-the-art multilingual word embeddings for over 30 languages based on fastText. fastText is a library for efficient learning of word representations and sentence classification. fastText can be used for making word embeddings using Skipgram, word2vec or CBOW (Continuous Bag of Words) and use it for text classification. For downloading the English (en) and Spanish (es) embeddings, you can use: # English fastText Wikipedia embeddings curl -Lo data/wiki.en.vec https://s3-us-west-1.amazonaws.com/fasttext-vectors/wiki.en.vec # Spanish fastText Wikipedia embeddings curl -Lo data/wiki.es.vec https://s3-us-west-1.amazonaws.com/fasttext-vectors/wiki.es.vec Facebook MUSE also has 110 large-scale high-quality, truth, bilingual dictionaries to ease the development and evaluation of cross-lingual word embeddings and multilingual NLP. These dictionaries are created using an internal translation tool. The dictionaries handle the polysemy (the coexistence of many possible meanings for a word) of words well. As mentioned earlier, MUSE has two ways to obtain cross-lingual word embeddings. The Supervised approach uses a training bilingual dictionary (or identical character strings as anchor points) to learn a mapping from the source to the target space using Procrustes alignment. To learn a mapping between the source and the target space, simply run: python supervised.py --src_lang en --tgt_lang es --src_emb data/wiki.en.vec --tgt_emb data/wiki.es.vec --n_iter 5 --dico_train default The unsupervised approach learns a mapping from the source to the target space using adversarial training and Procrustes refinement without any parallel data or anchor point. To learn a mapping using adversarial training and iterative Procrustes refinement, run: python unsupervised.py --src_lang en --tgt_lang es --src_emb data/wiki.en.vec --tgt_emb data/wiki.es.vec Facebook MUSE also has a simple script to evaluate the quality of monolingual or cross-lingual word embeddings on several tasks: Monolingual python evaluate.py --src_lang en --src_emb data/wiki.en.vec --max_vocab 200000 Cross-lingual python evaluate.py --src_lang en --tgt_lang es --src_emb data/wiki.en-es.en.vec --tgt_emb data/wiki.en-es.es.vec --max_vocab 200000 To know more about the functionalities of this library and to download other resources, you can go through the official GitHub repo here.

OpenAI has built a new deep neural network called MuseNet for composing music, the details of which it shared in a blog post yesterday. The research organization has made a prototype of MuseNet-powered co-composer available for users to try till May 12th. https://twitter.com/OpenAI/status/1121457782312460288 What is MuseNet? MuseNet uses the same general-purpose unsupervised technology as OpenAI’s GPT-2 language model, Sparse Transformer. This transformer allows MuseNet to predict the next note based on the given set of notes. To enable this behavior, Sparse Transformer uses something called “Sparse Attention”, where each of the output position computes weightings from a subset of input positions. For audio pieces, a 72-layer network with 24 attention heads is trained using the recompute and optimized kernels of Sparse Transformer. This provides the model long context that enables it to remember long term structure in a piece. For training the model, the researchers have collected training data from various sources. The dataset includes the MIDI files donated by ClassicalArchives and BitMidi. The dataset also includes data from online collections, including Jazz, Pop, African, Indian, and Arabic styles. The model is capable of generating 4-minute musical compositions with 10 different instruments and is aware of different music styles from composers like Bach, Mozart, Beatles, and more. It can also convincingly blend different music styles to create a completely new music piece. The MuseNet prototype, which is made available for users to try, only comes with a small subset of options. It supports two modes: In simple mode, users can listen to the uncurated samples generated by OpenAI. To generate a music piece yourself, you just need to choose a composer or style and an optional start of a famous piece. In advanced mode, users can directly interact with the model. Generating music in this mode will take much longer but will give an entirely new piece. Here’s how the advanced mode looks like: Source: OpenAI What are its limitations? The music generation tool is still a prototype so it does has some limitations: To generate each note, MuseNet calculates the probabilities across all possible notes and instruments. Though the model gives more priority to your instrument choices, there is a possibility that it will choose something else. MuseNet finds it difficult to generate a music piece in case of odd pairings of styles and instruments. The generated music will sound more natural if you pick instruments closest to the composer or band’s usual style. Many users have already started testing out the model. While some users are pretty impressed by the AI-generated music, some think that it is quite evident that the music is machine generated and lacks the emotional factor. Here’s an opinion shared by a Redditor for different music styles: “My take on the classical parts of it, as a classical pianist. Overall: stylistic coherency on the scale of ~15 seconds. Better than anything I've heard so far. Seems to have an attachment to pedal notes. Mozart: I would say Mozart's distinguishing characteristic as a composer is that every measure "sounds right". Even without knowing the piece, you can usually tell when a performer has made a mistake and deviated from the score. The Mozart samples sound... wrong. There are parallel 5ths everywhere. Bach: (I heard a bach sample in the live concert) - It had roughly the right consistency in the melody, but zero counterpoint, which is Bach's defining feature. Conditioning maybe not strong enough? Rachmaninoff: Known for lush musical textures and hauntingly beautiful melodies. The samples got the texture approximately right, although I would describe them more as murky more than lush. No melody to be heard.” Another user commented, “This may be academically interesting, but the music still sounds fake enough to be unpleasant (i.e. there's no way I'd spend any time listening to this voluntarily).” Though this model is in the early stages, an important question that comes in mind is who will own the generated music. “When discussing this with my friends, an interesting question came up: Who owns the music this produces? Couldn't one generate music and upload that to Spotify and get paid based off the number of listens?.” another user added. To know more in detail, visit the OpenAI’s official website. Also, check out an experimental concert by MuseNet that was live-streamed on Twitch. OpenAI researchers have developed Sparse Transformers, a neural network which can predict what comes next in a sequence OpenAI Five bots destroyed human Dota 2 players this weekend OpenAI Five beats pro Dota 2 players; wins 2-1 against the gamers

Netflix, the video-on-demand streaming company, has seen a surge in its growing number of users every day as well as in the viewership of its TV shows. It is constantly striving to provide an enriching experience to its viewers. To keep pace with the ever-increasing demands of user experience, Netflix is introducing a collection of tools and algorithms to make its content more audience relevant. AVA( Aesthetic Visual Analysis)- analyses large volumes of images obtained from video frames of a particular TV show to set as the title image for that show. Netflix understands that a more visually appealing title image plays an incredibly important role assisting a viewer find new shows and movies to watch. How title images are selected normally Usually, content editors had to go through tens of thousands of video frames for a show, to select a good title image. To give you a gist of the effort required-  a single one-hour episode of ‘Stranger Things’, consists of nearly 86,000 static video frames. Imagine sieving through each one of these frames painstakingly to find the perfect title image that will not only connect with the viewers, but also give them a gist of the storyline. To top it all up, the number of frames can go up to a million depending on the number of episodes in a show. This task of manually screening the frames is almost impossible and labor intensive, if not ineffective. Additionally, the editors choosing the image stills require an in-depth expertise of the source content that they’re intended to represent. Considering Netflix has an exponentially increasing catalog of shows, this presents a very challenging expectation for the editors to surface meaningful images from videos. Enter AVA, using its image classification algorithms for sorting the right image at the right time. What is AVA? The ever-growing number of images on the internet space has led to challenges in its processing and classification. To address this concern, a research team from University of Barcelona, Spain in collaboration with Xerox corporation has developed a method called Aesthetic Visual Analysis (AVA) as a research project. The project contains a vast database of over 2.5 lakh images combined with metadata such as aesthetic scores for images semantic labels for more than 60 classifications of images and many other characteristics. Using statistical concepts like standard deviation, mean score and variance, AVA rates images. Based on the distributions computed from these statistics, they assess the semantic challenges and choose the right images for the database. AVA primarily alleviates the issues of extensive benchmarking and trains more images. They also enable images to get a better aesthetic appeal. Computing performance can be significantly optimised to have lesser impact on the hardware. You can get more insights by reading the Research paper. The ‘AVA’ approach used at Netflix The process takes place in 3 steps: AVA starts by analysing images obtained through the process of frame annotation. This includes processing and annotating many different variables on every individual frame of video to best derive what the frame contains, and to understand its importance to the story. To keep up pace with the growing catalog of content, Netflix uses the Archer framework to process videos more efficiently. Archer splits the video into very tiny bits to aid parallel video processing. After the frames are obtained, they are subjected to a series of image recognition algorithms to build metadata. Metadata is further classified as visual, contextual and composition metadata.  To give you a brief overview- Visual Metadata: For brightness, sharpness and color Contextual Metadata: This is a combination of elements that are combined to derive meaning from the actions or movement of the actors, objects and camera in the frame. Eg: face detection, Motion estimation, Object Detection and camera shot identification Composition Metadata: For intricate image details based on core principles in photography, cinematography and visual aesthetic design such as depth of field and symmetry. Choosing the right Picture! The ‘best’ image is chosen considering three important aspects– the lead actors, visual range and sensitivity filters. Emphasis is given first to lead actors of the show since they make a visual impact. In order to identify the key character for a given episode, AVA utilizes a combination of face clustering and actor recognition to filter main characters from secondary characters or extras. The next thing, is the diversity of the images present in the video frames which includes camera positions, image details such as brightness, color, contrast to name a few. Keeping these in mind, image frames are easy to group based on similarities. This helps in developing image support vectors. The vectors primarily assist in designing an image diversity index where all the relevant images collected for an episode or even a movie can be scored based on visual appeal. Sensitive factors such as violence, nudity and advertisements are filtered and are allotted low priority in the image vectors. This way they are screened out completely in the process. Source: Netflix Blog What's in this for Netflix and its users? Netflix’s decision to use AVA will not only save manual labour, but also reduce the cost involved in having manpower source through millions of images in order to get that one perfect shot. This unique approach will help in obtaining meaningful images from video and thus enable creative teams to invest time in designing stunning artwork . As for its users, a good title image means establishing a deeper connection to the show’s characters and storyline, thus improving their overall experience. To understand the intricate  workings of AVA, you can read Netflix engineering team’s original post on this topic on Medium. How everyone at Netflix uses Jupyter notebooks from data scientists, machine learning engineers, to data analysts Netflix releases FlameScope Netflix bring in Verna Myers as new VP of Inclusion strategy to boost cultural diversity